基于机器学习的海上多层砂岩油藏精细注水开发实践

针对海上多层水驱砂岩油田作业成本高、小层测试数据少所导致的产、吸状况不清的问题，提出一种可同时学习多种井况条件的小层产、吸剖面预测模型。首先综合考虑影响小层产、吸状况的静态地质条件和动态开发特征，筛选并构造出主控因素，建立样本数据库。然后构建了巧妙的循环将神经网络算法和智能优化算法进行融合，内层循环以反向传播(back propagation, BP)神经网络为模型框架，遍历所有井样本，实现多维主控因素与产、吸剖面的机器学习；中层循环以量子进化算法为优化手段，实现神经网络内部权重和阈值自动优化；外层循环以测试误差为控制条件，保证模型的可靠性与最优化。最后将产、吸剖面预测模型应用于渤海P油田，分别对73口油井和84口水井的样本数据进行交叉验证，结果表明模型的平均测试误差仅为6.60%、4.36%。示例井组经分层调配等措施的综合治理之后，实现了井组日增油63 m³/d,综合含水率下降6%。该研究成果对老油田的精细注水和优化调整具有一定的指导意义。

Practice of fine water-flooding in offshore multilayer sandstone reservoirs based on machine learning

To address the problem of unclear production and suction conditions in small formations caused by high operating costs and low PLT test data in offshore multi-layer water-driven sandstone fields, a small formation production and suction profile prediction model that can simultaneously learn multiple well conditions is proposed. First, the static geological conditions and dynamic development characteristics that affect the production and suction profiles of small formations were considered, the main control factors were screened and constructed, and a sample database was established. Then a clever loop was constructed to integrate the neural network algorithm and the intelligent optimization algorithm. In the inner circulation, the BP neural network was used as the model framework to traverse all well samples and realize the machine learning of multi-dimensional controlling factors and production and suction profiles. In the middle circulation, the quantum evolution algorithm was used as the optimization method to realize the automatic optimization of the internal weight and threshold of the neural network. In the outer circulation, the test error was used as the control condition to ensure the reliability and optimization of the model. Finally, the production and suction profile prediction model was applied to the Bohai P oil field, and the sample data of 73 oil wells and 84 water wells were cross-validated respectively, and the results showed that the average testing error of the model was only 6.60% and 4.36%.The example well group achieved a daily oil increase of 60 ton per day and a comprehensive water cut reduction of 6% after comprehensive treatment of the sample well group by stratified allocation and other measures. The research results have certain guiding significance for fine water injection and optimization adjustment of old oilfield.